A dynamic multi-scale window continuous blood pressure prediction method and system based on PPG signal quality driving

Abstract

The application discloses a dynamic multi-scale window continuous blood pressure prediction method and system based on PPG signal quality driving, and relates to the technical fields of PPG signal processing and blood pressure prediction. The method comprises the following steps: acquiring original PPG signals and IMU data to obtain beat-by-beat PPG sequences; calculating each beat SQI and converting the same into window effective information quantity; updating the prediction window length, segment weight and memory coefficient according to the signal quality and the effective information quantity; inputting the effective window into a multi-scale time sequence pyramid encoder to extract and fuse features; outputting blood pressure prediction values through a trend constraint prediction decoder; and completing drift compensation and prompting through a calibration and output module. The system changes the prediction strategy according to the current wearing and motion state of a user, realizes fast response when the signal is good, increases effective information accumulation when the signal is poor, and keeps physiological continuity in prediction output. The application can be applied to continuous health monitoring scenes such as smart watches / wristbands, health monitoring wristbands and patch type monitoring devices.

Description

Technical Field

[0001] This invention relates to the fields of photoplethysmography (PPG) signal processing and artificial intelligence blood pressure prediction technology, specifically to a blood pressure prediction method and system that achieves cuffless continuous blood pressure estimation by dynamically determining the prediction window and fusing multi-scale features. Background Technology

[0002] Existing cuffless continuous blood pressure estimation techniques mainly include: the pulse wave conduction time (PTT / PAT) method, which utilizes the time delay relationship between cardiac electrical activity and peripheral pulse waves. Feature extraction relies heavily on human experience and is sensitive to wearing status, motion artifacts, and individual differences, using handcrafted features from single PPG signals combined with traditional machine learning methods. Deep learning methods using single PPG signals directly learn blood pressure-related features from PPG waveforms or their time-frequency representations using CNNs, RNNs, LSTMs, Transformers, attention mechanisms, or hybrid networks. Unsupervised pre-training and transfer learning methods utilize large amounts of unlabeled PPG data for representation learning, followed by transfer to labeled blood pressure prediction tasks.

[0003] However, existing technologies generally use fixed-length windows. When the PPG signal quality is good, a fixed-length window can cause response lag and computational redundancy; when the PPG signal quality is poor, the effective physiological information within a fixed-length window is insufficient, leading to decreased prediction stability. Furthermore, existing SQI or quality assessment methods are mostly used for preprocessing screening, elimination, or simple noise reduction, without further involvement in window length adjustment, weight allocation within the window, memory decay of low-quality segments, and prediction confidence control. The processing methods for low-quality segments lack mechanisms for progressive utilization and forgetting based on quality and time freshness.

[0004] Meanwhile, existing technologies using fixed short windows focus more on local pulsation patterns and struggle to reliably represent minute-level trends such as postprandial blood pressure elevation, post-exercise recovery, and nocturnal blood pressure rhythms, which are crucial for continuous blood pressure management. Furthermore, existing solutions treat adjacent window predictions as independent tasks, easily leading to non-physiological jumps in predicted values ​​between close moments, reducing the reliability of continuous monitoring results. Multi-sensor solutions also have high hardware dependencies, hindering low-burden daily continuous monitoring. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic multi-scale window continuous blood pressure prediction method and system based on PPG signal quality to solve the problems mentioned in the background art.

[0006] The specific technical solution provided by this invention is as follows: A dynamic multi-scale window continuous blood pressure prediction method based on PPG signal quality, comprising the following steps: Raw PPG signals and IMU data were acquired. The raw PPG signals were preprocessed by bandpass filtering and pulse wave peak detection. The single-beat signal segments extracted after beat-by-beat segmentation were combined to form beat-by-beat PPG sequences.

[0007] The DTW algorithm is used to calculate the morphological similarity between the current pulsation and the sliding template. After FFT transformation, the main frequency band energy ratio and motion interference intensity are calculated. The comprehensive SQI is obtained by weighted fusion of four dimensions, and the comprehensive SQI is converted into window effective information.

[0008] Preferably, the morphological similarity between the current pulsation and the sliding template is calculated using the DTW algorithm as follows: a sliding median template matching strategy is adopted, and the specified time period is divided into multiple time periods; the median waveform of the pulsation marked as high quality in the previous time period is defined as the sliding template; the morphological similarity between the current pulsation and the sliding template is calculated using the DTW algorithm to obtain a normalized similarity between 0 and 1.

[0009] Preferably, the four dimensions include: waveform template matching degree, signal-to-noise ratio, correlation between adjacent pulsations, and acceleration amplitude factor.

[0010] The effective information content is calculated based on the integrated SQI, and the prediction window length, segment weights, and memory coefficients are updated using signal quality and effective information content to output the pulsation sequence within the determined effective window.

[0011] Preferably, the output of the defined effective window pulsation sequence includes: Define the effective information content of the window: calculate the effective information content of the beats in the current candidate window using the beat SQI, memory coefficient, and time interval from the current prediction time; Construct an adaptive window length rule: Set a baseline window length, calculate the mean SQI of all impulses within the current window in real time, compress the target window length according to the baseline window length when the mean SQI of impulses is greater than the set high threshold, expand the target window length according to the baseline window length when the mean SQI of impulses is less than the set low threshold, and maintain the target window length according to the baseline window length when the mean SQI of impulses is within the set high and low threshold range. Weighting segments within the allocation window: Divide the effective information window into K segments by time, convert SQI scores into quality scores, convert time differences into timeliness scores, and assign a joint quality-timeliness weight to each segment; An adaptive forgetting mechanism is constructed: a memory coefficient is maintained for each beat within the window, and the updated memory coefficient is obtained each time it is updated. When the memory coefficient is lower than a preset threshold, the weight of the beat is reduced or it is excluded from the calculation of effective information or the selection of representative beats. The adaptive window length rule also includes: calculating the actual window length used at the current moment using an exponential moving average method. Develop a window refresh strategy: Execute window refresh when any preset condition is met, determine the pulsation sequence within the effective window based on fragment weights and memory coefficients, output the pulsation sequence within the effective window, and dynamically adjust the length and content of the effective window based on real-time data.

[0012] The pulsation sequence within the effective window is input into a multi-scale temporal pyramid encoder to extract the final fused feature vector.

[0013] Preferably, the multi-scale temporal pyramid encoder includes: Pulsation-level encoder: Input a single PPG pulsation waveform, each pulsation is normalized to a fixed length through interpolation, and a lightweight one-dimensional convolutional neural network is used to extract the internal morphological features of the pulsation, and output the pulsation feature vector; Sub-window level encoder: It assembles the continuous pulsation feature vectors into a sequence and divides them into sub-windows according to the dynamic window. Each sub-window is modeled for long-term time series using a lightweight selective state-space model and outputs a sub-window representation vector. Full-window encoder: Using the latest sub-window representation as the query vector and all sub-window representations as keys and values, it retrieves the historical sub-window information most relevant to the current state through cross-attention and outputs a full-window global feature vector.

[0014] Preferably, the extraction of the final fused feature vector includes: performing cross-scale feature fusion on three scale feature vectors, which contain a full-window-level global feature vector, a sub-window representation vector, and a representative pulsation feature vector; and extracting the final fused feature vector containing micro-pulsation morphology, meso-contextual patterns, and macro-blood pressure trends.

[0015] Based on the current raw blood pressure prediction value, historical blood pressure prediction status, current signal quality index (SQI), and signal quality index (SQI), the expected blood pressure trend is generated, and the blood pressure prediction value is output by the trend-constrained prediction decoder.

[0016] Preferably, the output of the blood pressure prediction value by the trend-constrained prediction decoder includes: receiving the final fused feature vector, inputting the final fused feature vector into the fully connected prediction head using the blood pressure prediction head, and outputting the current original blood pressure prediction value and its corresponding confidence level; constructing a temporal consistency constraint layer, using a lightweight state-space model to maintain the historical blood pressure trend hidden state, at each prediction time, the state-space model receives the current original blood pressure prediction value, the current window signal quality features, the historical signal quality features, and the previous hidden state, updates the current hidden state, and outputs the trend expectation value based on the current hidden state; the current window signal quality features include at least: the current window average SQI and the window effective information content; the historical signal quality features include at least: the SQI mean or SQI change trend within a preset historical period; and outputting the final blood pressure prediction value by weighted fusion of the original blood pressure prediction value and the trend expectation value.

[0017] Preferably, the output of the blood pressure prediction value by the trend constraint prediction decoder further includes: training the model during the training phase using a multi-task joint loss function including MAE loss, trend loss, consistency loss and calibration loss; through the multi-task joint loss function, the trend constraint prediction decoder optimizes multiple objectives simultaneously during the training process.

[0018] Based on the confidence level of the SQI judgment result, a light calibration or alarm mechanism is triggered to perform periodic calibration, complete drift compensation and prompt.

[0019] On the other hand, the present invention also provides a dynamic multi-scale window continuous blood pressure prediction system based on PPG signal quality, which executes a dynamic multi-scale window continuous blood pressure prediction method based on PPG signal quality. The system includes: PPG photoelectric sensor: Installed on wearable devices or other monitoring carriers, used to continuously acquire PPG waveform signals; Inertial Measurement Unit (IMU): Used to collect triaxial acceleration or triaxial angular velocity data, determine the user's motion state, and assist in motion artifact suppression; Signal preprocessing module: Performs bandpass filtering, baseline drift removal, motion artifact suppression, and beat-by-beat segmentation on the raw PPG waveform signal, and outputs a normalized beat sequence and its timestamp; Signal quality assessment engine: Calculates multi-dimensional quality indicators for each PPG waveform signal and merges them into a continuous signal quality index (SQI); Effective information window management engine: used to receive SQI sequence and pulsation time information, calculate the effective information of the window, dynamically adjust the current prediction window length, allocate the weight of sub-segments within the window, and perform adaptive forgetting and refreshing for persistent low-quality pulsations; Multi-scale temporal pyramid encoder: used to receive the pulsation sequence within the effective window determined by the effective information window management engine, extract multi-scale blood pressure related features through pulsation level, sub-window level and full window level encoding, and perform cross-scale fusion based on the weights output by the effective information window management engine; Trend-constrained prediction decoder: It receives the fused feature vector, combines it with the historical blood pressure trend, and outputs the current predicted values ​​of SBP, DBP, and MAP, as well as the prediction confidence. Calibration and Output Module: Used to perform initial personalized calibration, periodic light calibration, long-term drift compensation, and abnormal trend alerts. Outputs include SBP, DBP, MAP, confidence level, signal quality status, and prompts suggesting whether to remeasure or recalibrate.

[0020] Compared with the prior art, the beneficial effects achieved by the present invention are: (1) This invention uses SQI as a control variable that runs through window management, weight allocation, memory decay and confidence output, instead of just for preprocessing screening, thus realizing closed-loop adaptive signal quality driven by signal quality.

[0021] (2) In this invention, the window length is no longer fixed to a certain time period, but is dynamically determined by combining signal quality, time freshness and memory coefficient. This can take into account both the response speed in high-quality scenarios and the information accumulation in low-quality scenarios, and realize the management of prediction windows with effective information.

[0022] (3) The present invention uses a three-level time-series pyramid to extract pulsatile micromorphology, sub-window meso-context and full-window macro trend at the same time, which can capture short-term fluctuations and minute-level changes at the same time, making the multi-scale feature coverage more complete.

[0023] (4) Through the joint weighting of quality and timeliness and the adaptive forgetting mechanism, the system can gradually reduce the impact of low-quality fragments, avoid information breakage caused by simple discarding, and make the utilization of low-quality fragments smoother.

[0024] (5) The trend constraint prediction decoder used in this invention combines historical blood pressure trend status with current and historical SQI quality characteristics to generate trend expectation value, and adaptively fuses it with the original prediction result, which can reduce non-physiological jumps in adjacent predictions and make continuous prediction results physiologically reasonable.

[0025] (6) This invention does not rely on ECG electrodes. Conventional IMU is only used as a motion reference, which can reduce the hardware complexity, power consumption and wearing burden of wearable devices, and is suitable for low power deployment of single PPG. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the architecture of the dynamic multi-scale window continuous blood pressure prediction system provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the operation flow of the dynamic multi-scale window continuous blood pressure prediction method provided in the embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0028] The overall technical solution of this invention is mainly applied to continuous health monitoring scenarios in smartwatches, smart bracelets, health monitoring wristbands, patch-type monitoring devices, and bedside devices. These devices can continuously collect reflected or transmitted light signals from the user's skin via a PPG sensor and simultaneously collect motion status information via an IMU. Due to differences in the user's wearing status, movement amplitude, and sensor contact quality during rest, walking, running, sleep, and daily activities, the PPG waveform quality exhibits significant time-varying characteristics. Furthermore, blood pressure changes include both beat-by-beat waveform information and trend information over minute or longer time scales. Based on this, this invention proposes a dynamic multi-scale window mechanism driven by PPG signal quality, enabling the system to automatically adjust predictive behavior according to currently available physiological information.

[0029] Specifically, the present invention addresses the following technical problems based on the following embodiments: The problem with dynamic window technology is to enable the blood pressure prediction window to no longer be limited to a fixed time length, but to automatically expand and contract according to the PPG signal quality and the amount of effective information within the window. Under high-quality signals, the window should be shortened to improve the response speed, and under low-quality signals, the window should be extended to accumulate sufficient effective physiological information.

[0030] The problem of quality-driven information utilization: to further transform the signal quality evaluation results per beat (PPG) into window length control, segment weight allocation, memory decay, and output confidence, rather than simply using them for filtering.

[0031] The technical challenge of multi-scale feature fusion is to simultaneously extract pulsatile micromorphology, sub-window meso-context, and full-window macro-trend within a dynamically changing window, and stably fuse information from different scales.

[0032] Prediction consistency issues: Introduce historical trend states when continuously outputting SBP, DBP, and MAP, so that adjacent prediction results conform to the physiological law of continuous and gradual changes in human blood pressure, and reduce non-physiological jumps.

[0033] Single-sensor simplification technology challenge: To realize a continuous blood pressure prediction method and system suitable for wearable deployment, relying solely on a single PPG physiological signal and supplemented by a conventional IMU motion reference.

[0034] Example 1: Combination Figure 1 As shown in the figure, the dynamic multi-scale window continuous blood pressure prediction system based on PPG signal quality described in this embodiment includes the following functional units in its system architecture: PPG photoelectric sensor: installed on wearable devices or other monitoring carriers, used to continuously acquire PPG waveform data at a sampling rate of greater than or equal to 100Hz. In this embodiment, the sensor serves as the main physiological signal source of the system.

[0035] Inertial Measurement Unit (IMU): Synchronously acquires triaxial acceleration and / or triaxial angular velocity data to determine the user's motion state, assist in motion artifact suppression, and provide a basis for the motion intensity factor in signal quality evaluation.

[0036] The signal preprocessing module performs bandpass filtering, baseline drift removal, motion artifact suppression, and beat-by-beat segmentation on the raw PPG waveform signal, outputting a normalized beat sequence and its timestamp. In this embodiment, this module only performs basic signal processing and does not handle the final window decision.

[0037] Signal quality assessment engine: Calculates multi-dimensional quality indicators for each PPG waveform signal and merges them into a continuous signal quality index (SQI). In this embodiment, the SQI value is not only used to judge the quality of the signal, but also serves as a common driving variable for subsequent window management, weight allocation, and confidence output.

[0038] The effective information content window management engine receives SQI sequences and pulsation time information, calculates the effective information content of the window, dynamically adjusts the current prediction window length, assigns weights to segments within the window, and performs adaptive forgetting and refreshing for persistent low-quality pulsations. In this invention, "effective information content" is not simply the duration of the pulsation, nor is it simply discarding low-quality signals. Instead, it is a measure of usable information content after considering the pulsation memory coefficient, pulsation quality, and time-related decay. Window length adjustment is based on the SQI mean and effective information content, and a smooth transition is achieved through exponential moving averages. This ensures that low-quality signals are not abruptly deleted, but rather their contribution is gradually reduced according to their quality and duration; high-quality but earlier signals are balanced with the latest but lower-quality signals through a quality-time joint weighting.

[0039] Multi-scale temporal pyramid encoder: It is used to receive the pulsation sequence within the effective window determined by the effective information window management engine, extract multi-scale blood pressure related features through pulsation level, sub-window level and full window level encoding, and perform cross-scale fusion based on the weights output by the effective information window management engine. This avoids trend loss caused by using only a short window and also avoids slow response caused by simply using a long window.

[0040] Trend-constrained predictive decoder: It receives the fused feature vector, combines it with the historical blood pressure trend, outputs the current SBP, DBP, MAP predicted values ​​and prediction confidence, and reduces non-physiological jumps through adaptive trend constraints.

[0041] Calibration and Output Module: Used to perform initial personalized calibration, periodic light calibration, long-term drift compensation, and abnormal trend alerts. Output results may include SBP, DBP, MAP, confidence level, signal quality status, and prompts on whether to recommend remeasurement or recalibration.

[0042] Based on the aforementioned functional module architecture, the system of this invention forms a data flow pipeline from signal acquisition to blood pressure prediction output. The data flow between modules is as follows: PPG photoelectric sensor and inertial measurement unit (IMU) → signal preprocessing module → signal quality assessment engine → effective information window management engine → multi-scale temporal pyramid encoder → trend-constrained predictive decoder → calibration and output module → SBP / DBP / MAP predicted values ​​and confidence level output. Specifically, the effective information window management engine converts PPG signal quality into effective information and window control strategies; the multi-scale temporal pyramid encoder performs multi-scale feature encoding within a dynamic window; and the trend-constrained predictive decoder combines historical trend states to output blood pressure prediction values ​​that conform to physiological continuity.

[0043] Example 2 In this embodiment, the present invention, based on the system of Embodiment 1 which includes a specific functional module architecture, executes a dynamic multi-scale window continuous blood pressure prediction method driven by PPG signal quality. In this method, SQI is not an isolated quality label, but a control variable continuously transmitted between the signal quality assessment engine, the effective information window management engine, the multi-scale temporal pyramid encoder, and the trend constraint prediction decoder in the system. Based on this, the present invention can change the prediction strategy according to the user's current wearing and exercise status, achieving the overall effect of "rapid response when the signal is good, increased accumulation of effective information when the signal is poor, and maintaining physiological continuity in the prediction output".

[0044] Specifically, in combination Figure 2 As shown, the dynamic multi-scale window continuous blood pressure prediction method includes the following steps: Step S1: Acquire raw PPG and IMU data through PPG and IMU sensors, perform bandpass filtering and pulse wave peak detection preprocessing on the raw data, extract single beat signal segments after beat-by-beat segmentation, and combine the single beat signal segments to form beat-by-beat PPG sequence; Step S2: The DTW algorithm is used to calculate the morphological similarity between the current pulsation and the sliding template. After FFT transformation, the energy ratio of the main frequency band is calculated. The correlation between the current pulsation and the previous n pulsations is calculated simultaneously, and the motion interference intensity is calculated based on IMU data. The comprehensive SQI is obtained by weighted fusion of four dimensions, and the SQI is converted into window effective information. In this embodiment, a four-dimensional signal quality assessment model is constructed using a signal quality assessment engine. Four-dimensional indicators are calculated for each PPG waveform signal to comprehensively quantify the PPG signal quality from four dimensions: waveform morphology, frequency domain energy, timing stability, and motion interference. These four-dimensional indicators include: Waveform Template Matching (TQI): A sliding median template matching strategy is employed. The median waveform of several high-quality pulsations (10-20 high-quality pulsations selected in this embodiment) from the previous period is defined as the sliding template. The morphological similarity between the current pulsation and the sliding template is calculated using the Dynamic Time Warping (DTW) algorithm, resulting in a normalized similarity between 0 and 1. This index reflects the morphological consistency between the current pulsation and typical high-quality waveforms in this invention.

[0045] Signal-to-noise ratio (SNRI): This is calculated by performing a Fast Fourier Transform (FFT) on the current pulse, i.e., frequency domain analysis of the current pulse or a short segment, to calculate the ratio of energy within the 0.5-5Hz pulse main frequency band to the energy across the 0-10Hz full frequency band, and normalizing it to the 0-1 range. This metric measures the proportion of physiologically effective components to noise components in the PPG waveform signal. Compared to traditional time-domain SNRI calculations, frequency domain analysis can more accurately distinguish between physiological signals and noise, and is particularly suitable for assessing power line interference (50 / 60Hz) and high-frequency motion artifacts.

[0046] Adjacent pulse correlation coefficient (CI): Calculates the median Pearson correlation coefficient between the current pulse and the previous M (M can be 5 to 10) pulse waveforms. This index reflects the temporal stability of adjacent pulse waveforms. Sudden changes in waveform usually indicate a decrease in signal quality or an increase in motion artifacts. For example, CI < 0.4 usually suggests the presence of motion artifacts or arrhythmias and can be used as an auxiliary screening indicator for arrhythmias such as atrial fibrillation.

[0047] Acceleration Amplitude Factor (AMI): The triaxial acceleration amplitude or angular velocity amplitude corresponding to the current pulse time period is obtained from the inertial measurement unit (IMU) and normalized to 0 to 1, where the larger the value, the stronger the motion interference.

[0048] The SQI calculated by fusing the above four-dimensional indicators using a linear weighted fusion model is as follows:

[0049] Among them, weight Initial value can be set. The SQI value ranges from [0,1]. During model training or individualized calibration, the weights can be fine-tuned according to the data distribution. This invention comprehensively considers waveform morphology, frequency domain energy, temporal stability, and external motion disturbances through this fusion method, providing a unified quality input for subsequent dynamic window control.

[0050] Step S3: Calculate the effective information content based on SQI, update the prediction window length, segment weights, and memory coefficients using signal quality and effective information content, and output the determined pulsation sequence within the effective window.

[0051] In this embodiment, the effective information content window management engine transforms the PPG signal quality from an "evaluation result" into a "window control variable." The window in this invention is not determined solely by a fixed number of seconds, but rather by the effective information content window jointly determined by the duration, shot-by-shot quality index (SQI), time freshness, and memory factor. Specifically, it includes: Define the effective information content of a window: Let the current candidate window be... Contains several pulsations SQI is The memory coefficient is The time interval from the current prediction time is The effective information content of the window Represented as:

[0052] in, Indicates the first The normalized signal quality of each beat is important; higher quality beats contribute more to the amount of effective information. For the first The signal quality index of each segment. The memory coefficient of the old beats over time. As the memory coefficient falls below a threshold, the weight of the pulsation is gradually reduced or eliminated. The time decay coefficient, Indicates the first The time decay factor of a beat is such that the closer the beat is to the current moment, the closer its time decay factor is to 1, and the greater its contribution; the farther the beat is from the current moment, the closer its time decay factor is to 0, and the smaller its contribution. This definition makes the availability of a window depend not only on the window duration, but also on the quality of information within the window, whether historical fragments have decayed, and the distance of the fragments from the current moment.

[0053] Constructing adaptive window length rules: Setting a baseline window length The time is 5 minutes, configurable from 3 to 10 minutes; it calculates the average SQI of all beats within the current window in real time. When the mean Greater than the set high threshold At that time, the target window length is compressed to the baseline window length. 0.7 times that of the average value, to improve response speed in high-quality signal scenarios; when the average value is... Less than the set low threshold When the target window length is expanded to the baseline window length, the target window length is then increased. 1.5 times that of the average, to accumulate more effective physiological information; when the mean When the target window length is between the two thresholds, it remains near the baseline window. The window length is limited by... and For example, from 1 minute to 15 minutes. An exponential moving average is used to smoothly transition the actual window length used at the current moment. :

[0054] in The smoothing coefficient is set to 0.2 to 0.5 in this embodiment. The ideal window length is calculated based on factors such as current signal quality and motion intensity. When the signal quality is good... Larger values ​​are used to improve calculation accuracy when the signal quality is poor. Smaller size, used to improve response speed. The weight is the actual window length used in the previous time step. Determined by the smoothing coefficient, the smaller the smoothing coefficient, the greater the impact of the historical window length. This mechanism avoids frequent jumps in window length caused by short-term SQI fluctuations.

[0055] Weighting segments within the allocation window: Divide the effective information window into K segments by time. By converting SQI scores into quality scores and time differences into timeliness scores, assign a joint quality-timeliness weight to each segment. :

[0056]

[0057]

[0058] in, For the first Quality score function for each sub-segment For the first The timeliness score function for each sub-segment In this embodiment, the scaling factor for the quality score is... and As a learnable or configurable parameter, this mechanism allows the model to automatically weigh segments that are "better quality but older" against segments that are "newer but lower quality." Specifically, it uses an exponential function to map the SQI score to positive real numbers; as the SQI score increases, the quality score grows exponentially, emphasizing the importance of high-quality segments. The growth rate of the quality score is controlled, with larger values ​​indicating higher weight for quality. An exponential decay function is also used to map the time difference to a real number between 0 and 1. When the time difference is small, the timeliness score is close to 1, indicating good timeliness for that segment; when the time difference is large, the timeliness score decays exponentially, indicating poor timeliness for that segment. Control the decay rate of the timeliness score; the larger the value, the higher the weight of timeliness.

[0059] Constructing an adaptive forgetting mechanism: Maintaining a memory coefficient for each pulse within the window. The initial value is set to 1.0; the updated memory coefficient is obtained with each update. :

[0060] in, The memory coefficient before the update. This is the decay rate coefficient. When the pulsation quality is good ( When it is close to 1), Approaching 0, the memory coefficient decays slowly and remains at a high value; when the pulsation quality is poor ( When it is close to 0, As the memory coefficient approaches 1, it decays rapidly, and its value decreases quickly; when the memory coefficient... Below the preset threshold At this time, the pulsation is reduced in weight or excluded from the calculation of effective information or the selection of representative pulsations. The mechanism set in this invention is smoother than hard thresholding, which can reduce the continuous contamination caused by low-quality segments, while retaining some information of low-quality pulsations, providing a high-quality data source for multi-scale coding, and improving the accuracy and stability of the analysis.

[0061] A window refresh strategy is implemented: Window refresh is performed when any of the following conditions are met: the current window length reaches the adaptive target length or the effective information content reaches a threshold; the IMU detects a significant change in the user's motion state; consecutive low SQI segments accumulate for more than a preset duration; or the prediction confidence remains below a threshold. During refresh, 30% to 70% of the effective overlapping pulsations with the window before refresh are retained to maintain trend continuity and introduce new state information. Finally, the pulsation sequence within the effective window is determined based on segment weights and memory coefficients, and the pulsation sequence within the effective window is output. The length and content of the effective window are dynamically adjusted based on real-time data.

[0062] Step S4: Provide quality weights for the pulsation-level CNN, and input the pulsation sequence within the effective window into the multi-scale temporal pyramid encoder to extract fused features.

[0063] In this embodiment, the pulsation sequence within the valid window determined by the valid information window management engine is encoded based on the following encoder: Pulsation-level encoder: Input a single PPG pulsation waveform, and each pulsation is interpolated and normalized to a fixed length, such as 128 sampling points. This embodiment preferably uses a lightweight one-dimensional convolutional neural network to extract the internal morphological features of the pulsation and output a pulsation feature vector. This scale primarily characterizes microscopic information such as the slope of the ascending branch, peak width, waveform area, peak-valley relationship, and diphtheria notch.

[0064] Sub-window level encoder: The continuous pulsation feature vectors are arranged into a sequence and divided into N sub-windows according to the dynamic window output by S130. Each sub-window is modeled for long-range time series using a lightweight selective state-space model (Selective SSM), outputting a sub-window representation vector s_j∈R^d. SSM is chosen because it has linear complexity for long sequences, making it suitable for minute-level PPG sequences and edge deployments. Without changing the core concept of this invention, the sub-window level encoder can also employ other long-sequence models with linear or near-linear complexity, such as temporal convolutional networks, linear attention models, or other state-space models.

[0065] Full-window encoder: The full-window encoder uses the latest child window representation as the query vector and all child window representations as keys and values. It retrieves the most relevant historical child window information to the current state through cross-attention and outputs a full-window global feature vector. This design is used to characterize macroscopic trends and state changes throughout the entire effective window.

[0066] Cross-scale feature fusion: Features from three scales are fused in the following way:

[0067] in, Presentation layer normalization operation, Represents the global feature vector at the full window level. This represents vector concatenation. For child window indexing, For pulsation index, The weights of each child window, assigned based on the effective information content, are output by the effective information content window management engine. The sub-window represents the vector. The representative pulsation weights selected from each sub-window, This represents the representative pulsation feature vector. The final fused feature vector... It also includes microscopic pulsation patterns, mesoscopic contextual patterns, and macroscopic blood pressure trends.

[0068] Step S5: Generate trend expectation based on the current raw blood pressure prediction value, historical blood pressure trend status, and current and historical SQI quality characteristics, constrain the output blood pressure prediction value of the prediction decoder, and provide a quality threshold for outlier detection.

[0069] In this embodiment, a fused feature vector from a multi-scale temporal pyramid encoder is received. Then, the blood pressure prediction head is used to fuse the feature vectors. Input a fully connected prediction head, output the current raw blood pressure prediction value and its corresponding confidence level:

[0070]

[0071] in, This is the original predicted blood pressure value. This is the predicted systolic blood pressure value. This is the predicted value for diastolic blood pressure. This is the predicted value for mean arterial pressure. To indicate whether a remeasurement, a reduction in display frequency, or a calibration trigger is required. To predict the confidence level of systolic blood pressure, To predict the confidence level of diastolic blood pressure, Confidence level for predicting mean arterial pressure.

[0072] A temporal consistency constraint layer is constructed: the original predicted values ​​reflect the direct characteristics of the PPG signal within the current window, but may contain non-physiological jumps. Therefore, a lightweight state-space model is used to maintain the hidden state of historical blood pressure trends. Let the current input vector be... Based on the current raw blood pressure prediction value and mass state vector Formed by splicing, among which This includes at least the current window's average SQI, the window's effective information content, the SQI change, and the duration of continuous low quality. Specifically, the state-space model updates the current hidden state using the following formula:

[0073]

[0074] in, The hidden state represents the trend at the current moment. Let A be the hidden state from the previous time step, A be the historical state transition matrix, and B be the current input mapping matrix. For state bias, Using a hyperbolic tangent activation function, the current hidden state is mapped to the expected trend values ​​of systolic blood pressure, diastolic blood pressure, and mean arterial pressure through a trend output layer:

[0075] in, This represents the expected value of the trend at the current moment. and These are the weight matrix and bias vector of the trend output layer, which is a fully connected linear mapping layer. In this embodiment, a multilayer perceptron containing at least one fully connected layer and a nonlinear activation function can also be used. The final blood pressure prediction value is... The calculation is as follows:

[0076] in, The expected value of the trend. The adaptive weights for the current original predicted values ​​range from 0 to 1. Based on the mass state vector Calculated using the Sigmoid function:

[0077] in, and These are either trainable or calibrated parameters. When the average SQI and effective information content are high in the current window, improve... To enhance the response to the current raw prediction; when the current signal quality is low, the SQI continues to decline, or the effective information is insufficient, reduce This is to improve the expected value of the trend. If the SQI is below the minimum quality threshold or the amount of effective information is consistently insufficient, the output confidence level is reduced or the output of blood pressure values ​​is paused, and the user is prompted to wear the device again, remeasure, or perform calibration.

[0078] Multi-task joint optimization of loss function: This invention employs a multi-task joint loss function, including MAE loss, trend loss, consistency loss, and calibration loss, to train the model during the training phase.

[0079] Wherein, λ1, λ2, and λ3 are auxiliary loss weights, and in this embodiment, the recommended initial values ​​are 0.3, 0.2, and 0.1, respectively. For the total loss function, For the average absolute error loss, The trend loss function is... For consistency loss function, To calibrate the loss function.

[0080] Using mean absolute error loss To constrain the prediction errors of systolic blood pressure (SBP) and diastolic blood pressure (DBP) and improve the model's prediction accuracy, a trend loss function is used. The predicted blood pressure change direction is constrained to be consistent with the reference direction to ensure that the prediction results conform to the physiological laws of blood pressure changes and avoid non-physiological blood pressure change directions. A consistency loss function is used. Penalizes drastic changes in adjacent predicted values ​​that exceed physiologically reasonable ranges, ensuring physiological consistency of continuous prediction results and reducing non-physiological jumps. A calibration loss function is used. Constrain the matching relationship between prediction confidence and actual error to ensure that the confidence accurately reflects the reliability of the prediction results and improve the credibility of the system.

[0081] This invention, through the above multi-task joint loss function, enables the trend-constrained prediction decoder to simultaneously optimize multiple objectives during training, thereby improving the model's prediction accuracy, generalization ability, and physiological consistency, and providing users with more reliable blood pressure monitoring information.

[0082] Step S6: Based on the SQI judgment result confidence level, the calibration and output module triggers a light calibration or alarm mechanism to realize drift compensation and prompt, and provides a quality threshold for sliding into the next moment.

[0083] Example 3 In this embodiment, the present invention recommends the following settings for the key technical parameters involved:

[0084] An example of everyday wear scenarios is continuous blood pressure monitoring with a smartwatch.

[0085] In everyday wear scenarios, continuous blood pressure monitoring using a smartwatch includes: when the user wears a smartwatch with a built-in PPG sensor and IMU, the device continuously collects wrist PPG signals at a sampling rate of 100Hz or higher, and simultaneously records movement status, and operates as follows at different time periods: During sedentary office hours: PPG signal quality is high, for example, SQI is about 0.85. Based on the high SQI and high effective information content, the system of this invention compresses the window to about 0.7 times the reference window to improve the prediction refresh speed so as to track the trend of slow rise in postprandial blood pressure in a timely manner.

[0086] During walking commutes: SQI may be at a moderate level, for example, about 0.60. The system of this invention maintains a window length close to the baseline and assigns weights based on the SQI and timeliness differences of each segment, so that segments with smaller arm swings and more stable waveforms receive higher contributions.

[0087] During brief running periods: The SQI may drop to a low level, for example, around 0.35. After detecting the low-quality state, the system expands the window and applies memory decay to persistent low-quality fluctuations; the trend constraint layer simultaneously reduces the risk of anomalous jumps. If the continuous low SQI exceeds a preset duration, a window refresh is triggered and the output confidence is reduced.

[0088] The device can prompt users to perform a light calibration of the cuff blood pressure every 7 to 14 days, or trigger a calibration prompt in advance when the confidence level is consistently low, the wearing status changes significantly, or the predicted trend is abnormal.

[0089] For example, a nighttime sleep scenario—passive and imperceptible monitoring.

[0090] When a user wears the monitoring device during sleep, the wrist is relatively still at night, resulting in generally high PPG signal quality, such as an SQI of approximately 0.80 to 0.90. The system of this invention can maintain a short and stable effective window and output blood pressure prediction values ​​at preset intervals. The trend-constrained prediction decoder continuously maintains historical states throughout the night, thus smoothly tracking natural blood pressure fluctuations associated with different sleep stages such as light sleep, deep sleep, and REM sleep. When a rise in blood pressure during the night is detected, or other abnormal trends are observed, the system can output a trend alert. This scenario demonstrates the invention's rapid response and long-term trend tracking capabilities under high-quality signal conditions.

[0091] Therefore, given sufficient training data, reference cuff or invasive blood pressure data, and a reasonable calibration procedure, the system and method of this invention are expected to achieve the following technical effects: Prediction accuracy: Improved single-PPG blood pressure estimation accuracy through effective information window and multi-scale feature fusion. Specific compliance details regarding AAMI or BHS Grade A standards should be presented based on subsequent experimental data.

[0092] Motion robustness: In common motion scenarios such as walking and jogging, low-quality signals are no longer simply processed with equal weights. The system improves prediction stability through window expansion, quality weighting, and forgetting mechanisms.

[0093] Trend tracking capability: Minute-level dynamic windows can better capture slow-changing trends such as postprandial blood pressure elevation, post-exercise recovery, and nocturnal blood pressure rhythm.

[0094] Edge feasibility: The preferred lightweight CNN and selective SSM have low parameter count and linear complexity, making them suitable for deployment on wearable devices equipped with MCUs, DSPs or NPUs.

[0095] Power consumption and hardware advantages: Compared with the PPG plus ECG dual sensor solution, this invention uses a single PPG as the main physiological signal source, which can reduce hardware complexity and sensor-related power consumption.

[0096] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic multi-scale window continuous blood pressure prediction method based on PPG signal quality, characterized in that: The following steps are included: Raw PPG signals and IMU data were acquired. The raw PPG signals were preprocessed by bandpass filtering and pulse wave peak detection. The single beat signal segments extracted after beat-by-beat segmentation were combined to form beat-by-beat PPG sequences. The DTW algorithm is used to calculate the morphological similarity between the current pulsation and the sliding template. After FFT transformation, the main frequency band energy ratio and motion interference intensity are calculated. The comprehensive SQI is obtained by weighted fusion of four dimensions, and the comprehensive SQI is converted into window effective information content. The effective information content is calculated based on the integrated SQI, and the prediction window length, segment weights, and memory coefficients are updated using signal quality and effective information content to output the pulsation sequence within the determined effective window. The pulsation sequence within the effective window is input into a multi-scale temporal pyramid encoder to extract the final fused feature vector. Based on the current raw blood pressure prediction value, historical blood pressure prediction status, current signal quality index (SQI), and signal quality index (SQI), the expected blood pressure trend is generated, and the blood pressure prediction value is output by the trend-constrained prediction decoder. Based on the confidence level of the SQI judgment result, a light calibration or alarm mechanism is triggered to perform periodic calibration, complete drift compensation and prompt.

2. The dynamic multi-scale window continuous blood pressure prediction method based on PPG signal quality driven according to claim 1, characterized in that: The DTW algorithm is used to calculate the morphological similarity between the current pulsation and the sliding template, including: A sliding median template matching strategy is adopted, and the specified time period is divided into multiple time periods; Define the median waveform of the high-quality beats in the previous time period as the sliding template; The morphological similarity between the current pulsation and the sliding template is calculated using the DTW algorithm, resulting in a normalized similarity between 0 and 1.

3. The dynamic multi-scale window continuous blood pressure prediction method based on PPG signal quality driven according to claim 2, characterized in that: The four dimensions include: waveform template matching degree, signal-to-noise ratio, correlation between adjacent pulsations, and acceleration amplitude factor.

4. The dynamic multi-scale window continuous blood pressure prediction method based on PPG signal quality driven according to claim 3, characterized in that: The output of the defined effective window of pulsation sequence includes: Define the effective information content of the window: calculate the effective information content of the beats in the current candidate window using the beat SQI, memory coefficient, and time interval from the current prediction time; Construct an adaptive window length rule: Set a baseline window length, calculate the mean SQI of all impulses within the current window in real time, compress the target window length according to the baseline window length when the mean SQI of impulses is greater than the set high threshold, expand the target window length according to the baseline window length when the mean SQI of impulses is less than the set low threshold, and maintain the target window length according to the baseline window length when the mean SQI of impulses is within the set high and low threshold range. Weighting segments within the allocation window: Divide the effective information window into K segments by time, convert SQI scores into quality scores, convert time differences into timeliness scores, and assign a joint quality-timeliness weight to each segment; Construct an adaptive forgetting mechanism: Maintain a memory coefficient for each beat within the window, obtain the updated memory coefficient each time it is updated, and reduce or exclude the beat from the calculation of effective information or the selection of representative beats when the memory coefficient is lower than the preset threshold. Develop a window refresh strategy: Execute window refresh when any preset condition is met, determine the pulsation sequence within the effective window based on fragment weights and memory coefficients, output the pulsation sequence within the effective window, and dynamically adjust the length and content of the effective window based on real-time data.

5. The dynamic multi-scale window continuous blood pressure prediction method based on PPG signal quality driven according to claim 4, characterized in that: The construction of adaptive window length rules also includes: calculating the actual window length used at the current moment using an exponential moving average method.

6. The dynamic multi-scale window continuous blood pressure prediction method based on PPG signal quality driven according to claim 5, characterized in that: The multi-scale temporal pyramid encoder includes: Pulsation-level encoder: Input a single PPG pulsation waveform, each pulsation is normalized to a fixed length through interpolation, and a lightweight one-dimensional convolutional neural network is used to extract the internal morphological features of the pulsation, and output the pulsation feature vector; Sub-window level encoder: It assembles the continuous pulsation feature vectors into a sequence and divides them into sub-windows according to the dynamic window. Each sub-window is modeled for long-term time series using a lightweight selective state-space model and outputs a sub-window representation vector. Full-window encoder: Using the latest sub-window representation as the query vector and all sub-window representations as keys and values, it retrieves the historical sub-window information most relevant to the current state through cross-attention and outputs a full-window global feature vector.

7. The dynamic multi-scale window continuous blood pressure prediction method based on PPG signal quality driven according to claim 6, characterized in that: Extracting the final fused feature vector includes: Cross-scale feature fusion is performed on three scale feature vectors, which include a full-window-level global feature vector, a sub-window representation vector, and a representative pulsation feature vector. Extract the final fused feature vector that includes micro-pulsation morphology, meso-contextual patterns, and macro-blood pressure trends.

8. The dynamic multi-scale window continuous blood pressure prediction method based on PPG signal quality as described in claim 7, characterized in that: The blood pressure prediction values ​​output by the trend-constrained predictor decoder include: Receive the final fused feature vector, input the final fused feature vector into the fully connected prediction head using the blood pressure prediction head, and output the current original blood pressure prediction value and the corresponding confidence level; A temporal consistency constraint layer is constructed, and a lightweight state-space model is used to maintain the hidden state of historical blood pressure trends. At each prediction time, the state-space model receives the current original blood pressure prediction value, the current window signal quality characteristics, the historical signal quality characteristics, and the previous hidden state, updates the current hidden state, and outputs the expected value of the trend based on the current hidden state. The current window signal quality characteristics include at least: the current window average SQI and the window effective information content; Historical signal quality characteristics include at least: the average SQI or the SQI trend within a preset historical period; The final blood pressure prediction value is output by weighted fusion of the original blood pressure prediction value and the expected trend value.

9. The dynamic multi-scale window continuous blood pressure prediction method based on PPG signal quality driven according to claim 8, characterized in that: The output of the blood pressure prediction value by the trend constraint prediction decoder also includes: training the model during the training phase using a multi-task joint loss function that includes MAE loss, trend loss, consistency loss and calibration loss; through the multi-task joint loss function, the trend constraint prediction decoder optimizes multiple objectives simultaneously during training.

10. A dynamic multi-scale window continuous blood pressure prediction system based on PPG signal quality, executing the dynamic multi-scale window continuous blood pressure prediction method based on PPG signal quality as described in any one of claims 1-9, characterized in that: The system includes: PPG photoelectric sensor: Installed on wearable devices or other monitoring carriers, used to continuously acquire PPG waveform signals; Inertial Measurement Unit (IMU): Used to collect triaxial acceleration or triaxial angular velocity data, determine the user's motion state, and assist in motion artifact suppression; Signal preprocessing module: Performs bandpass filtering, baseline drift removal, motion artifact suppression, and beat-by-beat segmentation on the raw PPG waveform signal, and outputs a normalized beat sequence and its timestamp; Signal quality assessment engine: Calculates multi-dimensional quality indicators for each PPG waveform signal and merges them into a continuous signal quality index (SQI); Effective information window management engine: used to receive SQI sequence and pulsation time information, calculate the effective information of the window, dynamically adjust the current prediction window length, allocate the weight of sub-segments within the window, and perform adaptive forgetting and refreshing for persistent low-quality pulsations; Multi-scale temporal pyramid encoder: used to receive the pulsation sequence within the effective window determined by the effective information window management engine, extract multi-scale blood pressure related features through pulsation level, sub-window level and full window level encoding, and perform cross-scale fusion based on the weights output by the effective information window management engine; Trend-constrained prediction decoder: It receives the fused feature vector, combines it with the historical blood pressure trend, and outputs the current predicted values ​​of SBP, DBP, and MAP, as well as the prediction confidence. Calibration and Output Module: Used to perform initial personalized calibration, periodic light calibration, long-term drift compensation, and abnormal trend alerts. Outputs include SBP, DBP, MAP, confidence level, signal quality status, and prompts suggesting whether to remeasure or recalibrate.

Images